That reminds me of the article „a neural network in 11 lines of code“ from 2015 (although not on MNIST, rather some kind of XOR)

https://iamtrask.github.io/2015/07/12/basic-python-network/

Beautiful! Same in spirit as what I was going for, but even better. I really like the way the weights are updated in this one. Thanks for showing me this!

Tinygrad offers a superior MNIST implementation with minimal dependencies[0].

[0] https://github.com/tinygrad/tinygrad/blob/master/docs/mnist....

Autograd hides the backward pass in that implementation, but this code spells it out explicitly.

The description: "no dependencies."

The first line: "import numpy as np"

you got me, lol. to be fair the only reason numpy is there is because truly pure python doesn't have matmuls, and I didn't think reimplementing it wouldn't serve a didactic purpose.

A cool idea for a v2 without numpy would be implementing matmuls with lists, matrix transpose with zip() and switching np.exp for math.exp. And getting all that on fewer lines as possible.